ECE448/CS440: Introduction to Artificial Intelligence   Spring 2009

                       PROBLEM SET 4

Handed Out: 4/9/2009                                    Due:4/23/2009

       Programming the MultiLayer Perceptron

In this machine problem, you will be implementing the Back-propagation
algorithm for training a multi-layer neural network.  A discussion of
the algorithm is found in section 11.3 of the textbook.

       The Data Set

The data you will be using is from a real data set of mushrooms from
the Audubon Society Field Guide.  The data set is posted on the course
web site.  There are two files, a README file that describes the data
and the actual DATASET.  The data comprises 8124 instances of mushrooms
each having 22 attributes or features.  The first one is the target
value for the training data in the next 21 values.  The target values
are binary, either "p" for poisonous or "e" for edible.  Your network
is to be trained to reproduce these values as closely as possible.
The features are all discrete variables that take on only a finite number
of values.  You must transform these variables into a set of binary variables
valued only 0 or 1 which should be regarded as numerical rather than
logical values.  Use the following conversion method.  If a feature
can take k values, convert it into a sequence of k binary values so
that the lth bit is 1 if the feature is the lth symbol.  All other bits
for the feature are set to 0.  If a feature is missing, set all the bits
to 0.  In this way, each mushroom will correspond to a binary vector of 60
bits.

       The network

The input layer will have exactly 60 units.  Use one hidden layer with
as many units in it as you see fit.  You may try different values.  The
output layer will have two units, one for "poisonous" and one for "edible".
The decision rule is to label the sample according to which of the two
output units has the greater activation level.  Use the standard sigmoid
function as the unit non-linearity.

       Training

Assign randomly selected values for all the weights.  Use about 75%
of the data as training and the remaining 25% as testing data.  The
samples you hold out for testing is up to you.  Use back propagation
to determine the weights with a convergence criterion and a step size
of your choosing.

       Testing

Put the held out data into the network and use the decision rule above
to make the classification.  Compare the decision to the true category
given in the data set and compute the error rate.  Try various types
of networks and step sizes to get the best performance.

       Turn in the following

a) A diagram of your final best network
b) The performance of the networks you tried as a function of
  the number of hidden units, the step size, and the convergence
  criterion.
c) Your code